Driving · 2026-10-20

PaceVLA: Adaptive Temporal Action Quantization in Vision-Language-Action Models

Hojin Jeon, Jimin Lee, Taekwan Kim, Nayoung Oh, Taeyoung Kim, Myungkyu Koo, Jinmo Koo, Joonwoo Ahn, Jinwoo Shin

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

PaceVLA is an adaptive vision-language-action framework for temporally efficient robot execution. It uses a shared flow-matching action expert with mode-specific decoders for multiple temporal granularities and prediction horizons, plus a quantizability-guided mode selection (quantization judge + context-conditioned router) to choose fine vs quantized actions per inference step.

Why it matters

The authors target a speed–accuracy trade-off in temporal action quantization for VLAs: coarse execution can speed up free-space motion but may discard critical details in precision-sensitive phases. PaceVLA adapts temporal granularity and prediction horizon to what the context requires, aiming to improve execution efficiency while maintaining task success across simulation and real-world embodiments.

Method

Uses a shared flow-matching VLA action expert with mode-specific output decoders supporting multiple temporal action quantization levels and prediction horizons.
Trains a quantization judge to select fine-grained vs quantized action using LVLM-generated quantizability supervision, then uses a router to pick a specific mode within the eligible group.
Optimizes prediction and routing jointly (including quantizability regression and balancing losses) and adjusts controller tracking/gripper gains to support accelerated execution.

Limitation

The authors report that “most failures come from small targeting errors … or from the hands and tools blocking the contact region in the egocentric view,” and note failures where “missing information is in the observation, not in the trajectory,” suggesting limitations from perception quality (e.g., “better sensing, such as depth or tactile input, would add to the gains”).

Abstract (from arXiv)

Fast and reliable manipulation is crucial for practical robot deployment. Although Vision-Language-Action (VLA) models have recently shown remarkable progress as generalist robot policies, their execution often remains temporally inefficient. A key source of this inefficiency is that policies learn fine-grained temporal patterns from teleoperated demonstrations, requiring many low-level action steps. Temporal action quantization offers a simple way to accelerate execution by aggregating consecutive low-level actions into temporally coarser actions. However, applying action quantization without considering when fine-grained control is required introduces a speed-accuracy trade-off: accelerated actions may efficiently reduce redundant steps in some phases, while discarding critical motion details in precision-sensitive interactions. To address this, we propose PaceVLA, a unified framework for adaptive temporal execution in VLAs that integrates two coupled components: (1) mode-specific action decoders built on a shared action expert, supporting multiple temporal granularities and prediction horizons, and (2) a quantizability-guided mode selection scheme which decides whether coarse execution is appropriate, and a context-conditioned router that selects an execution mode at each inference step. Experiments across three multi-task simulation benchmarks and six real-world tasks show that PaceVLA improves execution efficiency while maintaining task success rate across multiple embodiments, including long-horizon and dexterous manipulation.

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-10-19 · Indexed by Pub-AI: 2026-10-20 UTC

preprint